Automated Segmentation of Organ Chambers Using Deep Learning Methods from Medical Imaging

1. A method of detecting whether or not a body chamber has an abnormal structure or function comprising: (a) providing a stack of images comprising, at least a representation of the body chamber inside the patient, as input to a system, (b) detecting the body chamber from each of the images using deep convolutional networks trained to locate the body chamber, (c) inferring a shape of the body chamber using a stacked auto-encoder (AE) network trained to delineate the body chamber, wherein the AE network segments the body chamber, (d) incorporating the inferred shape into a deformable model for segmentation, and (e) detecting whether or not the body chamber has an abnormal structure, wherein an abnormal structure is indicated by a body chamber clinical indicia that is different from a corresponding known standard clinical indicia for the body chamber.

2. The method according to claim 1 , wherein a structure of the deformable model of the body chamber is processed spatially and temporally to determine if function of the body chamber is abnormal.

3. The method according to claim 2 , further comprising quantifying a degree of abnormality of the body chamber.

4. The method according to claim 1 , further comprising performing contour alignment to reduce misalignment between multiple slices of medical images.

5. The method according to claim 1 , wherein the clinical indicia is selected from the group consisting of: a volume of the body chamber, an ejection fraction, a mass of the body chamber or a chamber's wall thickness of the body chamber.

6. The method according to claim 1 , wherein the body chamber is a chamber of a heart.

7. The method according to claim 6 , wherein the chamber of the heart is selected from the group consisting of a left ventricle, a right ventricle, a left atrium and a right atrium.

8. The method according to claim 1 , wherein the images comprise magnetic resonance imaging (MRI) images, ultrasound images, or CT scan data.

9. The method according to claim 1 , wherein the system is configured to utilize a training data set to initialize filters randomly to train the deep convolutional networks.

10. The method according to claim 9 , wherein the filters are convolved with the input medical images to obtain k convolved feature maps of size m 1 ×m 1 , computed as: Z l ⁡ [ i , j ] = ∑ k 1 = 1 a ⁢ ∑ k 2 = 1 a ⁢ F l ⁡ [ k 1 , k 2 ] ⁢ I ( i + k 1 - 1 , j + k 2 - 1 ] + b 0 ⁡ [ l ] , ( 1 ) for 1≤i,j≤m 1 , l=1, . . . , k, and m 1 =m−a+1; and wherein p×p non-overlapping regions in the convolved feature maps are computed as: P l ⁡ [ i 1 , j 1 ] = 1 p ⁢ ∑ i = 1 + ( i 1 - 1 ) ⁢ p i 1 ⁢ p ⁢ ∑ j = 1 + ( j 1 - 1 ) ⁢ p j 1 ⁢ p ⁢ C l ⁡ [ i , j ] , ( 2 ) for l≤i 1 , j 1 ≤m 2 , wherein m 2 =m 1 /p and p is chosen such that m 2 is an integer value.

11. The method according to claim 1 , further comprising aligning the images of the body chamber by performing contour alignment to reduce misalignment between the short-axis images.

13. The method according to claim 1 , further comprising identifying a segment of a body chamber from an output of a trained graph.

14. The method according to claim 1 , further comprising obtaining filters using a sparse autoencoder (AE), which acts as a pre-training step.

15. The method according to claim 13 , wherein the trained graph has two or more hidden layers.

16. The method according to claim 1 , wherein the AE network is a deep convolutional AE network.